RE: [PATCH] static_call,x86: Robustify trampoline patching

[David Laight]

From: Peter Zijlstra
> Sent: 03 November 2021 08:36
> 
> On Tue, Nov 02, 2021 at 05:20:05PM -0700, Andy Lutomirski wrote:
> > I think that's a big mistake -- any sane ENDBR-using scheme would
> > really prefer that ENDBR to be right next to the actual function body,
> > and really any scheme would benefit due to better cache locality.
> 
> Agreed, IBT/BTI want the landing pad in front of the actual function.
> 
> > But, more importantly, IMO any sane ENDBR-using scheme wants to
> > generate the indirect stub as part of code gen for the actual
> > function.
> 
> Sorta, I really want to be able to not have a landing pad for functions
> whose address is never taken. At that point it doesn't matter if it gets
> generated along with the function and then stripped/poisoned later, or
> generated later.
> 
> As such, the landing pad should not be part of the function proper,
> direct calls should never observe it.
> 
> Less landing pads is more better.

One problem is when a direct call is 'too far' for a call instruction.
IIRC this can happen in arm64 with modules (all 64bit except x86?).
So an indirect call has to be used instead - which needs the landing pad.
Although it may actually be better to put a trampoline (landing pad
+ near jump) elsewhere and have the module loader do the correct fixup.
(Is the loader already generating a trampoline in the module code?)
The function body can then be cache-line aligned - with its benefits.

Can't anything that can write instructions always use a retpoline
to implement a jump indirect to an arbitrary address?
(Not to mention just generating the required code rather than a call.)

AFAICT CFI is all about detecting invalid values in function pointer tables.
It doesn't really protect in any way from JIT code doing incorrect things.

	David

-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)